RL-TOPS: An Architecture for Modularity and Re-Use in Reinforcement Learning

This paper introduces the RL-TOPs architecturefor robot learning, a hybrid systemcombining teleo-reactive planning and reinforcementlearning techniques. The aim ofthis system is to speed up learning by decomposingcomplex tasks into hierarchies ofsimple behaviours which can be learnt moreeasily. Behaviours learnt in this way cansubsequently be re-used to solve a variety ofproblems, reducing the need to learn everynew task from scratch. It is even possibleto learn multiple...